GPU VulnDB

Database/Firmware, BMC & network fabric

Linux kernel efa driver: admin-queue completion context used after free on command error

CVSS 7.8CVE-2026-31493Firmware, BMC & network fabriccurated

Impact

When an efa admin command completes with an error, the driver prints data out of a completion context that the polling/interrupt handler already freed, so it reads memory that may have been reallocated for something else. efa is the RDMA transport on AWS GPU instances, where it carries all-reduce traffic between nodes in a training cluster. The practical consequences are kernel-log disclosure of whatever now occupies that memory and corruption in the fabric control path; a wedged RDMA device on one node stalls the whole collective. The admin queue is driven by the driver, not by tenants, so this is a reliability and info-leak bug rather than a tenant-reachable escalation.

Who can reach it

Local to the host. The efa admin queue is not exposed to tenant processes or to the network; reaching the bug requires an admin command to fail, which normally comes from device or driver state rather than from a caller.

What to do

Move to a stable kernel containing the fix (five stable branches have it, per the commits in the record). efa is a module, but unloading it tears down every RDMA queue pair on the node, so in practice this is a node drain followed by reboot rather than a live module swap. No vendor mitigation is described in the record.

References

Related entries

All Firmware, BMC & network fabric entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.